Evaluating the Impact of Medical Image Reconstruction on Downstream AI Fairness and Performance

Matteo Wohlrapp, Niklas Bubeck, Daniel Rueckert, William Lotter. Evaluating the Impact of Medical Image Reconstruction on Downstream AI Fairness and Performance. MIDL (2026).

---

Code Structure

The code is organized into the following directories, each containing its own README with detailed information:

Core Directories

• evaluation/ - Analysis and plot generation for the submission
• reconstruction_bias/ - Segmentation and classification for UCSF, U-Net training for CheXpert and UCSF from scratch, and U-Net mitigation methods
• sde/ - SDE training from scratch for UCSF and CheXpert
• sde_fairness/ - SDE fine-tuning for fairness on UCSF and CheXpert
• gan/ - GAN training from scratch for UCSF and CheXpert
• gan_fairness/ - GAN fine-tuning for fairness on UCSF and CheXpert

Implementation Notes

• U-Net implementation is our own custom implementation
• GAN and SDE implementations are forked and adapted, with separate codebases for training from scratch versus fairness mitigation
• CheXpert classification is performed using the torchxrayvision repository